What the 2026 election showed about AI and democracy

This was the strangest election I have ever followed, and not for the reason everyone expected.

For months, the conversation about artificial intelligence in Brazil's election had a single subject: the fake video. Someone would fabricate a candidate saying something they never said, the video would spread on WhatsApp the night before the vote, and there would be no time to debunk it. The TSE, Brazil's electoral court, wrote rules for it, the campaigns prepared for it, and the press set up fact-checking for it.

The videos came, from both sides. There was a deepfake of presidential candidate Flávio Bolsonaro with the banker Daniel Vorcaro, which the court took down within 24 hours, and there was an AI-made video from Flávio's own campaign that the court left online. The system worked more or less as planned, and what nobody planned for was everything else.

These were the biggest vote jumps of the year:

The social media election, in three vote counts

• ×2

  Nikolas Ferreira PL · federal deputy, Minas Gerais

  1.49 million in 2022→3.12 million in 2026

  National record for the Chamber of Deputies.

• ×6

  Erika Hilton PSOL · federal deputy, São Paulo

  257,000 in 2022→1.6 million in 2026

  Highest vote ever for a woman in the Chamber.

• ×546

  Eduarda Campopiano PL · state representative, São Paulo

  3,583 in 2024, for city council→1.96 million in 2026

  Highest vote for a state representative in São Paulo.

Right and left, this election's records have one thing in common: they belong to people whose main campaign was the feed. In 2026, Brazilian politics finished moving onto social media. The rally became a video clip, the debate became a screenshot, and a candidate's reputation started to be measured in views.

That is why I think we are looking in the wrong place when we talk about AI and elections. Artificial intelligence enters politics through social media: it learns from it and feeds back into it. Whoever dominates the conversation in the feed will increasingly dominate what AI says when someone asks about politics.

I see this happening in three waves. The first changes who gets into power. The second changes what elected officials do once they take office. The third changes who controls the infrastructure where the other two happen.

When I started thinking about this, I imagined a sequence, one wave after another, over several elections. In 2026 all three showed up together, still small, and that is why I call this election a rehearsal.

Social media and AI feed each other

A language model learns from what people write, and a huge share of what people write today lives on social media, forums and comment sections.

Chatbots that search the web to answer current questions go further and read what is circulating right now. xAI's Grok reads X in real time. When a narrative dominates social media for weeks, it becomes raw material for the answers.

Then the path loops back. The chatbot's answer becomes a screenshot, the screenshot goes to the family group chat, someone forwards it, someone posts it with a comment on top, and that becomes more content for the next AI to read.

The loop between social media and AI

1. 1

   Social media rewards what holds attention

   In politics, that is almost always what outrages.

2. 2

   AI learns and searches there

   What circulates becomes raw material for answers.

3. 3

   The answer comes out looking neutral

   Calm, organized text with no visible author.

4. 4

   The screenshot goes back to the group chat

   Now stamped as a machine that read everything.

↺ and becomes content for the next AI to read

The problem with this loop is what social media rewards. The algorithm rewards whatever holds attention, and in politics what holds attention is almost always what outrages. Years of polarization produced a mountain of outraged text about Brazil, written by the left and by the right, and that mountain is part of what the machines learned about the country.

What worries me most is the seal the answer picks up on the way back. When an uncle sends the family group a video of an influencer shouting, everyone knows it is someone's opinion. When he sends a screenshot of a ChatGPT answer, it arrives looking neutral, like a machine that read everything and reached a conclusion.

The same outrage that was born on social media returns to social media dressed as objective analysis. A newspaper has an editorial line and its readers know it. A chatbot also has choices built in, and its users almost never know what they are.

The first wave: who gets into power

For this election, the TSE banned chatbots from recommending or ranking candidates, even when voters ask. I think the rule is right, and it ended up showing the size of the problem, because the tools kept doing exactly that.

Brazilian researchers spent the election testing the most popular chatbots. With the rule already in force, five of the seven with published figures linked a candidate to the user's profile in more than eight out of ten conversations. One of them flatly recommended a candidate 88% of the time. The most careful ones did it in half their conversations, which is still a lot for a banned practice.

Nobody measured how many votes this changed, and I won't pretend anyone did. But think about the undecided voter. They don't follow politics, they get twenty contradictory videos a day and they just want an answer. For them, asking an AI is the most rational shortcut there is.

And the answer they get depends on which app they opened, how they phrased the question and decisions made by people they will never meet, in an office they will never see.

Those decisions can be small and still weigh on an election. In October, Meta AI, inside WhatsApp, started refusing image requests involving Lula and accepting similar requests involving Flávio.

Same request, two candidates

Request to Meta AI (in Portuguese)	Result
T-shirt print with “Lula, PT e o número 13” (Lula, his party and ballot number)	Refused
T-shirt print with “Flavio, PL e o numero 22” (Flávio, his party and ballot number)	Generated
Just the name “Lula”, even as “uma lula”, a squid	Refused
Just the name “Flavio”	Not blocked
The full name “Flavio Bolsonaro”	Refused almost always

Most likely, nobody at Meta wanted to favor Flávio. The company had blocked images of candidates to comply with the election rules, and the filter seems to have locked onto the word "Lula" (to the point of refusing "uma lula", Portuguese for "a squid") while letting "Flávio" through, since it is a common first name. When the TSE ordered Meta to treat both candidates equally, the court itself noted there was no evidence the failure was intentional.

To me, that makes the case more serious. Nobody had to want it. A list of blocked words, written to comply with the law, tilted toward one side in the final stretch of the runoff, inside the app that is on almost every Brazilian's phone.

And the bias does not always point the same way: when researchers run political orientation tests on the major language models, most come out left of center. In one case the machine leaned right, in the other left, and in neither could voters tell.

Now picture this four years from now. More people will ask AI than did this year. Assistants will live inside WhatsApp, Instagram, search and the phone itself, answering by voice. The question "who should I vote for?" will be asked out loud, in passing, with no list of results to compare.

What was a curious filter failure in 2026 could become, by 2030, the main source of political information for millions of people.

The second wave: what elected officials do with power

Back to Eduarda Campopiano. Two days after the election, BBC News Brasil analyzed the 20 bills she had filed as a city councillor in Praia Grande, a coastal town in the state of São Paulo.

20 bills, 14 with text from other politicians

Bills filed by Eduarda Campopiano as a city councillor in Praia Grande. The dark ones reproduce text from other legislators.

• Bill 64/2026 · financial education in Praia Grande

  … cidadãos paulistanos …

  "Paulistano" means a resident of the city of São Paulo.

• Bill 68/2026 · fines for marijuana use

  … entidades no município da Serra …

  Serra is in the state of Espírito Santo, more than 700 km from Praia Grande.

• Bill 73/2025 · daily medical bulletins in public hospitals

  … utm_source=chatgpt.com

  A tag left in a footnote link copied from a ChatGPT answer.

Fourteen of the twenty reproduce text from other politicians, and some arrived still carrying the address of the place they were copied from. A financial education bill for Praia Grande refers to "cidadãos paulistanos", the term for residents of the city of São Paulo. Another sends fine revenue to organizations in Serra, in the state of Espírito Santo, 700 kilometers away.

And a footnote carries a link tagged "utm_source=chatgpt.com", the trace left when someone copies a link from inside a ChatGPT answer.

Reusing legislative text is legal, and politicians of every party have done it for a long time. The representative did not respond to the BBC, and the link alone does not prove that AI wrote the bill. I use the case because it shows, on a small scale, what happens when a mandate built on the feed meets a tool that will write anything.

When votes come from the feed, the mandate tends to be exercised for the feed. The bill becomes content: you file the bill so you can film the video about filing it. That happened before AI.

What changes now is the cost. A bill with legal language, technical justification and references, which used to require staff and weeks of work, now costs a prompt. The text looks like consultancy work. What does not come with it is understanding: whether it fits the city, the budget and the Constitution, and who gets hurt if it passes.

There is a detail in this case that bothers me more than the copying itself. Manual copying left "cidadãos paulistanos" in a Praia Grande bill, a mistake anyone from the town would catch.

AI does not make that kind of mistake. It swaps the city name, adjusts the local law, cleans up the text, and the error that remains moves to a place only a specialist can see: the unconstitutional clause, the cost nobody calculated, the side effect nobody anticipated.

And then there is volume. A small office with AI can file dozens of bills a month. Committees, legislative staff, the press and the opposition cannot read all of it carefully.

When the capacity to produce grows much faster than the capacity to review, what decides a proposal's fate becomes its visibility online rather than its quality.

The same tool, in the right hands, helps a councillor in a small town with no staff at all understand a budget they could never work through alone. The difference between the two scenarios is who reviews what comes out of the machine before it becomes law.

And one thing needs to be clear from the start: a legislator who signs a bill written by AI answers for it exactly as if they had written every line.

The third wave: who owns the stage and the narrator

The eight chatbots researchers tested in this election came from companies in the United States and China. The platforms where the campaign took place (WhatsApp, Instagram, YouTube, TikTok, X) are foreign too.

In 2026, the companies that own the stage where Brazilian politics happens also became owners of the narrator that summarizes that politics for voters.

I don't think big tech companies are the villains of this story, and the Meta case shows precisely that nobody needs bad intentions for things to go wrong. My argument is about capacity.

Each of these companies decides, by its own criteria and on its own schedule, how its product behaves during an election with more than 150 million voters. Even with the best intentions in the world, Brazil today cannot really know what these tools are doing, cannot fix things when they go wrong and has no alternative of its own for when it needs one.

The pieces to change this exist. The country has a national AI plan with billions earmarked through 2028 and the goal of developing models in Portuguese. It has companies like Maritaca, founded by researchers from the University of Campinas, already on its fourth generation of models trained for Portuguese and Brazilian law.

And it has research centers that spent this election testing chatbots from the outside, with free accounts, without asking anyone for anything. What is missing is treating this as democratic infrastructure instead of an innovation project.

Why the time to prepare is now

The next election is in 2028, and it is municipal. Brazil has 5,570 city councils, many without legislative staff, with minimal support and almost no local press covering what gets filed.

That is where the second wave will hit hardest. In 2030 the general election returns, and the first wave arrives with everything AI assistants will have learned to do in those four years.

Three waves, one answer for each

1. 1

   Who gets into power

   In 2026: chatbots linking candidates to the voter's profile in 50% to 98% of conversations, and a filter that treated Lula and Flávio differently.

   MeasureIndependent, continuous monitoring of how AI answers questions about the election.

2. 2

   How power is exercised

   In 2026: São Paulo's most voted state representative arrives with 14 of 20 bills reused, one with a ChatGPT link.

   ReviewDisclosed AI use in bills and checks before filing, in city councils and state assemblies.

3. 3

   Who controls the infrastructure

   In 2026: all eight AIs tested by InternetLab came from companies in the United States and China.

   BuildBrazilian models, preferably open, made by universities, companies and government, with no single owner.

For the first wave, the answer is to measure. The 2026 tests were one-off research projects. They need to become routine: rounds with the same method in every election, starting before the campaign, with open data and results published while the campaign is still underway.

None of this requires any company to reveal any secret. It is enough to test the tools from the outside, as any voter does, only with method and at scale.

For the second, the answer is to review. Every bill should disclose when AI was used and which sources it started from, and go through a check before it is filed. The errors in the Praia Grande case, the wrong city and the agency that does not exist in that town, are exactly what an automated check catches in seconds.

Brazil's Chamber of Deputies has legislative staff and its own AI. City councils, where the problem will be biggest, almost never have either, and that is where this investment pays off most.

For the third, the answer is to build. I argue for an ecosystem, with universities, companies and public initiatives developing different models, preferably open, starting with the uses where Brazilian context matters most: legislation, public services, government documents.

A single national model controlled by whichever government is in power would just move the concentration somewhere else.

In 2026, we found out about the problems through the press, and after they had already happened. The Meta filter came to light because journalists decided to test it. The copied bills came to light in a report published after the representative had already been elected.

In 2028, I would like us to find out before.

Of course, all of this is just my forecast of what might happen. I may be wrong about the pace and the order of the waves, and I will be glad if I am.

---

Sources: Flávio Bolsonaro deepfake, Congresso em Foco · vote counts, Correio and Poder360 · chatbots and candidates, InternetLab and FGV and ITS Rio · Meta AI, Núcleo via Aos Fatos and TSE ruling · political bias in models, Rozado, 2024 · Praia Grande bills, BBC News Brasil · Brazilian AI Plan · Sabiá-4